UI-Vision: leaderboard

Given a desktop screenshot from one of 83 applications and a text query, the model must point at the named UI element; the score is the source's Final Avg of the Basic, Functional and Spatial settings (1,772, 1,772 and 1,935 queries per model). Unlike ScreenSpot-Pro it splits queries by how the element is described (name, function, or position relative to other elements).

Metric: Element grounding accuracy (%), Final Avg of the Basic, Functional and Spatial settings (point inside the target element's box). Source: raw.githubusercontent.com. Saturation forecast: Around 2029. 12 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Aguvis-7B13.7
2OS-Atlas-Base-7B9.02
3CogAgent-9B-202412208.94
4Claude 3.7 Sonnet (20250219)8.27#421
5ShowUI-2B5.94
6SeeClick5.39
7MiniCPM-V-2.64.34#3523
8Qwen 2 VL 7B Instruct2.7#2594
9InternVL2.5-8B2.09#2401
10Qwen 2.5 VL 7B Instruct0.85#1893
11Gemini 2.0 Flash (001)0.3#727
12InternVL2-8B0.09#3359

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=ui-vision · How It Works · Data refreshed daily, snapshot 2026-11-17.